libc/x86: ensure the stack 16-byte aligned when tasks created
Currently Renderscript sample code RsBalls crashed on x86 when SSE2
enabled. The root cause is that the stack was not 16-byte aligned
from the beginning when the processes/threads were created, so the
RsBalls crashed when SSE2 instructions tried to access the variables
on the stack.
- For the thread created by fork():
Its stack alignment is determined by crtbegin_{dynamic, static}.S
- For the thread created by pthread_create():
Its stack alignment is determined by clone.S. __thread_entry( ) is
a standard C function. In order to have its stack be aligned with
16 byte properly, __thread_entry() needs the stack with following
layout when it is called:
layout #1 (correct)
--------------
| |
-------------- <--ESP (ECX - 20)
| ret EIP |
-------------- <--ECX - 16
| arg0 |
-------------- <--ECX - 12
| arg1 |
-------------- <--ECX - 8
| arg2 |
-------------- <--ECX - 4
| unused |
-------------- <--ECX (16-byte boundary)
But it has following layout for now:
layout #2: (incorrect)
--------------
| |
-------------- <--ESP (ECX - 16)
| unused |
-------------- <--ECX - 12
| arg0 |
-------------- <--ECX - 8
| arg1 |
-------------- <--ECX - 4
| arg2 |
-------------- <--ECX (16-byte boundary)
Fixed in this patch.
Change-Id: Ibe01f64db14be14033c505d854c73033556ddaa8
Signed-off-by: Michael Liao <michael.liao@intel.com>
Signed-off-by: H.J. Lu <hongjiu.lu@intel.com>
Signed-off-by: Jack Ren <jack.ren@intel.com>
Signed-off-by: Bruce Beare <bruce.j.beare@intel.com>
diff --git a/libc/arch-x86/bionic/clone.S b/libc/arch-x86/bionic/clone.S
index 8abb7c8..352d23c 100644
--- a/libc/arch-x86/bionic/clone.S
+++ b/libc/arch-x86/bionic/clone.S
@@ -13,16 +13,21 @@
pushl %ebx
pushl %ecx
movl 16(%esp), %ecx
- movl 20(%esp), %ebx
+
+ # save tls
+ movl %ecx, %ebx
+ # 16-byte alignment on child stack
+ andl $~15, %ecx
# insert arguments onto the child stack
movl 12(%esp), %eax
- movl %eax, -12(%ecx)
+ movl %eax, -16(%ecx)
movl 24(%esp), %eax
- movl %eax, -8(%ecx)
- movl %ecx, -4(%ecx)
+ movl %eax, -12(%ecx)
+ movl %ebx, -8(%ecx)
subl $16, %ecx
+ movl 20(%esp), %ebx
movl $__NR_clone, %eax
int $0x80
test %eax, %eax
@@ -40,7 +45,7 @@
# we're in the child thread now, call __thread_entry
# with the appropriate arguments on the child stack
# we already placed most of them
- jmp __thread_entry
+ call __thread_entry
hlt
2: